Identifying and Characterizing Binding Sites and Assessing

Figure 1. SiteMap surface, site points, and cocrystallized ligand for 1ett, exterior of pocket. ... They can also be used to quickly evaluate virtual ...
0 downloads 0 Views 255KB Size
J. Chem. Inf. Model. 2009, 49, 377–389

377

Identifying and Characterizing Binding Sites and Assessing Druggability Thomas A. Halgren* Schro¨dinger, Inc., 120 West 45th Street, New York, New York 10036 Received September 8, 2008

Identification and characterization of binding sites is key in the process of structure-based drug design. In some cases there may not be any information about the binding site for a target of interest. In other cases, a putative binding site has been identified by computational or experimental means, but the druggability of the target is not known. Even when a site for a given target is known, it may be desirable to find additional sites whose targeting could produce a desired biological response. A new program, called SiteMap, is presented for identifying and analyzing binding sites and for predicting target druggability. In a large-scale validation, SiteMap correctly identifies the known binding site as the top-ranked site in 86% of the cases, with best results (>98%) coming for sites that bind ligands with subnanomolar affinity. In addition, a modified version of the score employed for binding-site identification allows SiteMap to accurately classify the druggability of proteins as measured by their ability to bind passively absorbed small molecules tightly. In characterizing binding sites, SiteMap provides quantitative and graphical information that can help guide efforts to critically assess virtual hits in a lead-discovery application or to modify ligand structure to enhance potency or improve physical properties in a lead-optimization context. INTRODUCTION

Understanding the structure and function of protein binding sites is a cornerstone of structure-based drug design. Developing this understanding requires knowledge of both the location and physical properties of the binding site. In addition, the identification of small-molecule binding sites as modulators of protein-protein interactions is of increasing interest. Furthermore, even when a validated binding site has been identified, it is often important to find additional potential binding sites where appropriate targeting could result in different biological effects or new classes of compounds. When the binding site is not known from a 3-D structure or from other experimental data, computational methods can be employed to suggest likely locations. Models that rely on geometric properties, such as POCKET,1 SURFNET,2 APROPOS,3 LIGSITE,4 CAST,5 PASS,6 and CASTp,7 have long been used for this purpose. These methods are generally very fast and are accurate when the binding sites are welldefined pockets. In developing their refined LIGSITEcsc model, Huang and Schroeder have shown that the degree of conservation of key residues in the putative binding pocket can be used to improve LIGSITE’s ability to detect drugbinding sites.8 PocketFinder, a method developed by Abagyan and co-workers, expands on geometric methods by contouring a smoothed van der Waals potential for the protein to identify candidate ligand-binding sites.9 This partially accounts for physical properties but neglects electrostatic and desolvation effects. Nayal and Honig have employed a large, comprehensive set of physiochemical, structural, and geometric descriptors in developing SCREEN,10 which has been shown to accurately identify binding sites in a training set * Corresponding author phone: (973)-744-0163; e-mail: halgren@ schrodinger.com.

of 99 cocrystallized complexes. Other methods have been developed that use fragment probes to find and characterize binding sites.11-13 These methods more accurately account for the physical properties of the putative binding sites but tend to be computationally expensive. Knowledge-based approaches have also been used with varying degrees of success.14,15 Finally, methods that combine physics and knowledge have been developed. For example, Coleman et al. developed a method that employs a local version of the MAPPOD approach recently proposed by Cheng et al.16 to identify likely binding sites.17 This method performed reasonably well for predictions on a limited set of targets. When combined with a statistical residue-coupling analysis, it also suggested a possible allosteric binding site in p38R MAP kinase that has some precedent in another kinase structure. In another example, Joughin et al. developed a method that combines the solvated electrostatic potential, surface curvature, and amino acid identity to predict phosphopeptide binding sites.18 When the location of the primary binding site is known, medicinal chemistry efforts to design better ligands can profit from a better understanding of the degree to which known ligands are, or fail to be, complementary to the receptor as well as from a critical assessment of the degree to which the occupancy of accessible but unexplored regions by appropriate ligand functionality can be expected to promote binding or could be used to improve the physical properties of the ligand without lessening its binding affinity. Such assessments can assist in the evaluation and optimization both of known binding molecules and of virtual screening hits. It is also important to understand the potential druggability of the site. It is estimated that 60% of small-molecule drugdiscovery projects fail because the target is found to not be druggable.19 Furthermore, it is estimated that only 10% of the proteins encoded in the human genome are druggable

10.1021/ci800324m CCC: $40.75  2009 American Chemical Society Published on Web 01/20/2009

378 J. Chem. Inf. Model., Vol. 49, No. 2, 2009

Figure 1. SiteMap surface, site points, and cocrystallized ligand for 1ett, exterior of pocket.

by oral small molecules.20 Given that most therapeutic projects in the pharmaceutical industry continue to pursue small-molecule, orally available therapeutics, the ability to accurately predict target druggability, before much time and money is spent on discovery efforts, is invaluable. Compared to binding-site identification, less effort has been placed on computational methods for predicting druggability, but the MAPPOD model of Cheng et al. has been shown to have high predictive ability.16 In this work we describe SiteMap,21,22 a program we have developed for identifying binding sites and for predicting their druggability. We describe in detail how SiteMap works and validate its utility across a large set of diverse proteins. We also illustrate its ability to characterize binding sites with quantitative and graphical descriptors. SiteMap is fast enough to be used routinely in drug-discovery studies. Thus, proteins with roughly 5000 atoms, including hydrogens, take about 2-3 min on a single CPU of a 2.4 GHz Intel Pentium 4 workstation, while proteins with roughly 8000 atoms typically take about 5 min and proteins with 12,000 atoms take about 9 min; for comparison, the average size of the proteins used in this work is 4250 atoms. These attributes make SiteMap a promising tool for binding-site analysis, virtualhit assessment, and lead optimization. RESULTS AND DISCUSSION

Site Maps. To illustrate the graphical feedback provided in a typical application, Figures 1 and 2 show the cocrystallized ligand for the thrombin 1ett receptor (Figure 3) and the “site points” for the binding site generated by SiteMap (white) in the context of the receptor structure and of the gray, translucent SiteMap surface (see the Methods section). The first focuses on relatively exposed regions of the site, while the second profiles the buried specificity pocket. Figures 4 and 5, taken from the same viewpoints, display the hydrophobic (yellow), hydrogen-bond donor (blue), and hydrogen-bond acceptor (red) maps but for clarity suppress the receptor surface. These figures show that portions of the hydrophobic groups of the ligand occupy hydrophobic regions and that the donors and acceptors of the ligand for the most part lie in or close to appropriate donor and acceptor

HALGREN

Figure 2. SiteMap surface and site points for 1ett, specificity pocket. The protein is shown in wire-frame.

Figure 3. Thrombin 1ett inhibitor 4-TAPAP.

Figure 4. Hydrophobic, donor, and acceptor maps for 1ett, exterior of pocket.

regions. On close examination, it is clear that SiteMap finds that some elements of the ligand structure, such as the nonhydrogen-bonded NH2 of the benzamidinium group, do not support the binding. In addition, the hydrogen-bonded ligand carbonyl group in Figures 4 and 5 just misses the red acceptor region. But SiteMap is correct on this score because the N···O distance, 3.33 Å, is too long for a strong hydrogen bond. Overall, however, the match is reasonably good; in many other cases, it is quite striking. In contrast to techniques that color-code the receptor surface, the maps and the properties generated by SiteMap depend on the site as a whole, not just on the character of the closest receptor atom. Moreover, the maps explicitly show the shape and suggest the extent of the regions

BINDING SITES

AND

DRUGGABILITY

J. Chem. Inf. Model., Vol. 49, No. 2, 2009 379 Table 1. Percent Success in Locating the Cocrystallized Site in 538 Proteins as a Function of the Ligand Binding Affinitya comparison all sites 1 mM best-scoring site is correctb largest site is correct binding site not found

85.9

98.5

87.6

81.2

65.4

78.1

89.6

80.0

72.4

65.4

7.8

1.5

5.5

11.8

26.8

a 67 sites have ligand binding affinities of